Active perception during angiogenesis: filopodia speed up Notch selection of tip cells in silico and in vivo

How do cells make efficient collective decisions during tissue morphogenesis? Humans and other organisms use feedback between movement and sensing known as ‘sensorimotor coordination’ or ‘active perception’ to inform behaviour, but active perception has not before been investigated at a cellular level within organs. Here we provide the first proof of concept in silico/in vivo study demonstrating that filopodia (actin-rich, dynamic, finger-like cell membrane protrusions) play an unexpected role in speeding up collective endothelial decisions during the time-constrained process of ‘tip cell’ selection during blood vessel formation (angiogenesis). We first validate simulation predictions in vivo with live imaging of zebrafish intersegmental vessel growth. Further simulation studies then indicate the effect is due to the coupled positive feedback between movement and sensing on filopodia conferring a bistable switch-like property to Notch lateral inhibition, ensuring tip selection is a rapid and robust process. We then employ measures from computational neuroscience to assess whether filopodia function as a primitive (basal) form of active perception and find evidence in support. By viewing cell behaviour through the ‘basal cognitive lens' we acquire a fresh perspective on the tip cell selection process, revealing a hidden, yet vital time-keeping role for filopodia. Finally, we discuss a myriad of new and exciting research directions stemming from our conceptual approach to interpreting cell behaviour.

This article is part of the theme issue ‘Basal cognition: multicellularity, neurons and the cognitive lens’.

Morphogenesis of multicellular tissues and organs is a remarkably precise, robust and adaptive process [1–3]. How do cells manage to coordinate their behaviour so precisely and rapidly, given the complexities and time constraints of the developing embryo environment they reside within? We will argue, and provide the first proof of concept evidence, that cells in tissue development have evolved a basal cognitive process of active perception in order to deliver on the adaptive problem-solving required of them during morphogenesis.

We focus here on angiogenesis, the process of new blood vessel sprouting from pre-existing ones.

Angiogenesis is critical for tissue development, homeostasis and repair across organs in the body and is frequently dysregulated in disease [4–6]. Angiogenesis is driven first and foremost by sprouting endothelial cells (ECs) as they form the new blood vessel walls. EC behaviour is highly dynamic and needs to be tightly coordinated in space and time to effectively perfuse the surrounding tissue. Despite extensive progress to better understand angiogenesis since the discovery of its role in cancer progression in the 1970s [7,8], many aspects of the timing and cell dynamics involved are not yet understood, largely due to an early lack of live-imaging methods [9]. Vascular networks are highly specialized morphologically and functionally in different organs of the body [10,11]. Although a remarkably consistent core set of molecular pathways and processes have been identified that appear to lie at the heart of angiogenesis across tissues and species [12] organ-specific fine-tuned adaptations to this core are beginning to be uncovered [13,14]. It is clear that much remains to be understood about how ECs gain information about the local tissue in order to adapt and construct organ-specific vascular network topologies and functions [15–17].

In the first core step of angiogenesis, ECs in a vessel collectively decide (select) who will be the most migratory (the ‘tip’ cell) and lead a new vessel ‘sprout’ that branches out from the pre-existing vessel (figure 1a) the other cells then follow behind (stalk cells). Tip cells are well-studied and characterized by their long thin, highly dynamic protrusions—filopodia (figure 1b). Two sprouts eventually fuse to form vessel loops that can then support blood flow (figure 1a). This process repeats iteratively, building up a dense looped network (the ‘primary plexus’) that is later remodelled and pruned through cell rearrangement and blood flow to a more familiar hierarchical tree-like structure [18,19]. Figure 1. Mechanisms of tip selection during vascular network formation (angiogenesis): (a) a pre-existing vessel becomes activated by factors in a hypoxic tissue environment (low in oxygen) starting the process of tip selection where an alternating ‘salt and pepper’ pattern of tip (green) and stalk cells (purple) emerges. Tip cells then migrate out, extending new branch ‘sprouts’, which fuse (anastomosis) to form loops supporting blood flow. The speed of selection determines the branch spacing properties of the resulting blood vessel network. Fast patterning leads to a denser network, slow patterning leads to a sparser network. (b) Representative time-lapse images of filopodia extension from a tip cell migrating from the dorsal aorta (DA) upwards to form the intersegmental vessel (ISV) of a Tg(kdrl:ras-mCherry)s896 embryo. The dotted blue lines highlight the nucleus of the tip cell. Pink arrows mark filopodia. (c) The Vegf/Notch tip selection pathways. Vegf interacts with Vegfr-2, expressed at the surface of the endothelial cells and on filopodia.

Vegfr-2 activation triggers Dll4 gene expression up-regulation and actin polymerization causing filopodia extension (e.g. via Pi3 K). In turn, Dll4 ligands activate Notch receptors on the neighbouring cell triggering cleavage of the Notch intracellular domain (NICD), which translocates to the nucleus reducing gene expression of Vegfr-2 receptors (and other pathways). Differences in Vegfr activation are amplified by this lateral inhibition competition between the cells until one cell has more Vegfr-2 receptors (which can continue stimulating the actin migratory response) and Dll4 (tip cell) inhibiting its neighbour (stalk cell, which has higher Notch activation levels and less Vegfr). (d) Two ways to conceptually frame how Notch decisions take place. One way corresponds to the prevailing view of the process ‘decide then move’ which designates filopodia as the end result of the selection decision-making process—prominent on migrating tip cells. The ‘move while deciding’ view, which takes into account the relative timescales of events, implicates filopodia ahead of Notch, driving the decision-making process. (Online version in colour.)

Crucially, we recently found through an in silico/in vivo study that if the cells are slower to collectively select the new tip cells, then this has real consequences for the resulting vascular network topology: the slower the selection, the less dense the network (figure 1a) [17]. We have begun to uncover a growing number of ‘temporal regulators’ of EC collective decisions (e.g. tissue-derived factors like semaphorins [17] and cell-specific factors such as the tetraspanin-like protein Tm4sf18 [20]) that can fine tune branch density in a potentially organ-specific manner.

Here we demonstrate that filopodia, rather than just being features of the selected migratory tip cells, provide a vital, overlooked ‘temporal regulator’ role, keeping the selection ‘decision-making’ process timely. This highlights a novel role for cell shape changes in enhancing cell decision-making. Namely, filopodia extensions and retractions (movements) alter receptor locations and activations (sensations), amplifying differences between the cells in a positive feedback cycle. This is a form of sensorimotor coordination which, we argue, could constitute a basal form of active perception (AP). To our knowledge, active perception has until now only been studied and identified in higher organisms and prokaryotic cells [21,22].

First, we will give a brief introduction and context to the biological and basal cognitive concepts proposed in this study. We detail the tip selection process and present our time-ordering argument for the overlooked role of filopodia during selection. Second, we present our original research findings providing: (a) the first in vivo evidence during zebrafish embryo development that filopodia loss slows selection as predicted in silico in our cell-based simulations with the previously well-validated memAgent-spring model (MSM) [23–27], (b) simulations identifying that filopodia enable rapid and robust selection by conferring bistable decision-making properties to individual ECs due precisely to the feedback coupling of sensation and movement of filopodia and (c) a simple information theory metric to extract quantitative ‘signatures’ of active perception, which has potential to be extended and applied to future in vitro/in vivo and in silico experiments aiming to detect and further delineate the role of cell-level AP in tissues.

Filopodia are well documented during angiogenesis in most organs in the body. We propose that their time-keeping role during selection could be a ubiquitous part of the angiogenic process. Moreover, AP properties could be endowed to other types of cells by different processes and shape-changing mechanisms. The scope of this study, however, is to provide the first proof of concept and simple, small step. We therefore conclude with a discussion into the potentially wide-reaching new research avenues this study opens up, both drilling down into the several different detailed molecular mechanisms potentially underlying AP in cells of different tissue-beds and ways that AP could present a novel ‘higher-level’ target to therapeutically modify cell behaviour to tune branching to a desired density in diseased or engineered tissues.

The process to select new tip cells is currently considered to be well-understood and already well-characterized to occur via Notch-Dll4 lateral inhibition [28–30] triggered primarily by Vascular Endothelial Growth Factor (Vegf) released when the surrounding tissue becomes hypoxic (low in oxygen; figure 1c) [31,32]. However, this widely accepted selection mechanism does not consider timing or the specific dynamics of the process. We recently argued that this has caused us to miss a critical, speed-controlling feature, required for normal selection [9,33]. The multiple cycles of gene expression required by Notch-Dll4 lateral inhibition to amplify initially small differences in input Vegf signal of neighbouring ECs on a vessel to select the heterogenous pattern of tip cells is seemingly incompatible with the rapid, dynamic changes in EC state, identity and behaviour observed in angiogenesis [34,35]. We recently showed that selection of tip cells is a specifically time-constrained step of the process during zebrafish intersegmental vessel formation [20]. In this phase of zebrafish development, tip cells emerge within a ‘selection window’ of approximately 8 h ending at 24 h post fertilization (hpf), after which no new tip cells are selected. One cycle of Notch-Dll4 lateral inhibition takes approximately 4–6 h in vitro in mouse ECs [36] suggesting the currently accepted selection process would have to take closer to 16–24 h to select each tip cell. This is true even accounting for the more rapid gene expression cycles in zebrafish (approx. 2 h for one cycle, [37]). Thus, we hypothesize that Notch lateral inhibition alone is too slow to account for in vivo selection.

So how do endothelial cells make rapid tip selection decisions? Early simulations with the MSM of Vegf–Notch selection first pointed to filopodia as an unexpected temporal regulator of the process [24], however the significance of those early simulations was not clear until now. Given that filopodia are normally used as a marker of the selected tip cell phenotype, this was a surprising prediction—how could the consequence of selection be part of the cause? We recently presented a carefully revised time-ordering of the precise events during selection, which highlighted a missed role for filopodia before tip cells are selected [9].

Time-ordered account of selection: When Vegf binds and activates Vegf receptors (Vegfr) on existing endothelial cells, the cells respond on the order of seconds by locally polymerizing actin and pushing out the cell membrane into thin protrusive shapes—filopodia—which are then rapidly extended and retracted [38,39]. These in turn move the Vegf sensors through the environment, as they reside on the deforming cell membrane. This creates a rapid positive feedback to Vegf signalling, which in itself can amplify and structure the differential Vegf input between neighbouring cells, requiring fewer cycles through the slow (on the order of hours) Notch gene-regulatory lateral inhibition mechanism to ‘stabilize’ the decision. Thus, filopodia are chronologically primary in the decision-making process, and Notch is secondary—cells ‘move while deciding’ (figure 1c), contrary to the prevailing view that filopodia movement are associated with the selected tip cells ‘decide then move’ [40–42].

The role of filopodia in angiogenesis has been surprisingly elusive, despite their clear abundance during the process. They were first identified as a guidance mechanism in angiogenesis, helping the selected tip cells to follow Vegf gradients [43]. However, they were later deemed to be ‘dispensable for guidance’ in zebrafish ISV sprouting, albeit still slowing down sprouting when reduced [44]. Filopodia together with lamellipodia are thought to anchor tip cells to extracellular substrates by focal contact points, together pulling the tip cell forward during migration and sprouting [45] though filopodia are known to be inhibitory to migration in other cell migration contexts [46]. They play a key role in facilitating the formation of anastomoses with other ECs [25,47,48]. However, these filopodia likely play a number of roles on ECs, both before and after tip cell phenotypes are selected. Filopodia have previously been defined to expand the spatial limits of cell–cell communication during Notch lateral inhibition patterning in drosophila bristle selection [49], by allowing them to reach out and contact more distant cells, but have not to our knowledge before been investigated as a temporal adaptor of Notch patterning, altering the speed and dynamic properties of Notch lateral inhibition.

The chronological ordering of the selection process (figure 1d) brings movement to the forefront and highlights a potential role for ‘sensorimotor’ feedback in informing Notch selection decisions. ‘Moving while deciding’ and sensorimotor feedback are central to early definitions of AP in the field of psychology [50–53] where AP is well known to enhance human decision-making and execution of other cognitive tasks such as language acquisition [54]. Thus, we consider whether filopodia may constitute a simple cell-level form of AP enhancing collective tip selection decisions.

A particularly relevant example of the necessity of active perception at the human scale is found in the study of saccadic eye movements. In this context, sensory input obtained by a stationary/passive individual is not sufficient to facilitate its behaviour without local movements of sensors on the otherwise immobilized body. The necessity of movement is related to the anatomy of a human eye—which has a highly concentrated cluster of sensory cells at a region known as the fovea. The area of the fovea is small, and sensory acuity falls off rapidly in regions further away from it. Vision is largely facilitated by the eye undergoing abrupt movements, purposefully darting the fovea around visual scenes and gathering relevant information (at a rate comparable to an optimal Bayesian model). An observer would not be able to accurately characterize a visual scene without the ability to move their fovea [55].

Indeed, biology across scales is rife with examples where AP is required for behaviour including organisms ranging from Escherichia coli to Caenorrhabditis elegans, Drosophila and humans [21,55–57]. At the closest scale to ECs—single prokaryotic cells—the run and tumble behaviour of E. coli during chemotaxis is a well-established example of such a case, where the noise induced by the random arrival of chemoattractant molecules makes it physically impossible for a stationary bacterium to know which direction is ‘up’ the chemoattractant gradient [56,58–60]. Though their particular modes of movement may differ, the commonality between these examples is that the agents in question all must use coordinated sensorimotor activity to gather enough information to facilitate their behaviours.

We aim here, however, not just to make a simple analogy that filopodia on ECs are akin to active perception in higher or single-celled organisms, but to take the first steps in creating the analysis tools needed to rigorously prove that the effect of filopodia on EC selection speed is specifically due to the coupling of movement and sensation that they generate. As AP has become more widely recognized and studied in a diversity of subject areas, we can draw on a rich repertoire of tools and approaches from psychology [50–53], robotics [61–63] and neuroscience [64,65] for investigation of AP in cell biology.

There is an initiative among leading biologists to reframe the notion of what it means to explain the causes of a biological event [66]. It has been suggested that the classic feed-forward view based on molecular and genetic pathways, particularly in the context of developmental biology, falls short of explaining the adaptability and plasticity of large-scale anatomical pattern formation even when they accurately describe the interactions of the individual molecular components [9,67,68]. Recent proposals prescribe the basal cognitive lens [69]—a strategy based on exploiting the deep parallels between anatomical pattern regulation and aspects of cognition such as perception, memory and learning to better model and ultimately control these processes. Thus, this study fits neatly into the paradigm of basal cognition, framing the anatomical pattern formation in angiogenesis in terms of perception, action and behaviour.

Early simulation predictions with the memAgent model [24] predicted that filopodia and Vegf gradients alter selection speed. Significant additions and extensions were rapidly made to this primitive agent-based model to incorporate a spring mesh to better model cell shape changes of the cell membrane by the actin cortex beneath [25]. Thereafter, the model was referred to as the memAgent-spring model (MSM) and has been widely used and extended to explore many other roles of filopodia and cell–cell interaction mechanisms during angiogenesis, with a wealth of predictions validated in vivo [26,27,35,70]. Here, we present the first confirmation and extended study showing the MSM, with its more realistic and verified cell shape-changing spring mesh, recapitulates the precursor models first predictions. In order to answer the question ‘do filopodia impact Notch selection timing’, we constrain the MSM to simulate only the very first, small rapid protrusive filopodia movements of ECs that are otherwise fixed in space, mimicking a first pre-existing vessel stimulated by Vegf in a tissue (see Methods—MSM).

We found that even small changes to the parameter determining the probability of extending a filopodium (F) in the model (initiating a new one or continuing to extend an existing one) resulted in a delayed selection of the tip/stalk alternating ‘salt and pepper’ pattern (figure 2a). F was normally set to 2 in all published works with this model. To assess loss-of-function (LOF) and gain-of-function (GOF) effects, we varied this value in increments of 0.5 in the range [0 : 3] with the most distinct changes to selection timing found with the values 0.5 and 3 (figure 2b,c). To be clear, by LOF and GOF we mean partial rather than total reductions and increases in function. We define the ‘time to pattern’ as the time until certain criteria capturing the salt and pepper pattern are stably met (see Methods—MSM). In agreement with the precursor model, a reduction in VEGF gradient steepness or concentration also slow patterning (figure 2c and electronic supplementary material, figure S1). New simulations predict that GOF filopodia extension can rescue patterning speed in LOF VEGF conditions (figure 2c,d). This points to an intrinsic interrelation of filopodia effects on tip selection speed with the quality of local activating signal in their environment. Figure 2. Filopodia loss-of-function (LOF) slows down Notch selection. (a) Time–space plots showing simulations of the MSM with ten cells fixed in place in a pre-existing vessel, as they detect Vegf, extend/retract filopodia and select an alternating ‘salt and pepper’ pattern of tip/stalk cells via VEGF/Notch lateral inhibition. Frames shown every 100 time steps vertically (Dll4 level high-low shown green-purple). Cutting the probability of filopodia extension (F = 0.5 ‘reduced filopodia’) slows down the collective patterning.

In the simulations shown, only one tip is selected in twice the available time window compared to five with normal filopodial activity. (b,c) Violin plots showing that increasing filopodial activity (gain-of-function, GOF) speeds up patterning in silico both for normal and reduced VEGF gradient, and decreasing filopodial activity (LOF) slows down patterning. Blue shading corresponds to estimated probability density of timing values in that range. Upward arrows indicate pattern not achieved within the time window of the simulation for the corresponding parameter settings. Pink line indicates median value, black lines indicate middle quartiles, whiskers indicate upper and lower quartiles. N = 100 simulations ran for 2000 time steps. (d) A representative image of a cell in the dorsal aorta (DA), showing filopodia formation before tip cells have been selected to leave the DA into the ISV. The dotted blue line highlights the nucleus of the EC, pink arrows mark filopodia. (e) Time-lapse images of sprouting ISVs in control and latrunculin B (LatB)-treated Tg(kdrl:nlsEGFP)zf109 embryos from 19 hpf, and quantification of the number of ECs selected to sprout per ISV over time. Red brackets indicate dividing cells. Nuclei are pseudocoloured. Fewer tip cells are selected to sprout in LatB-treated embryos and the selection process is slower. Scale bar, 25 µm. N = 2 independent experiments (55 ISV tip cells were quantified from 15 embryos). (f) Lateral views of pErk immunostaining (red) in ECs (green) of WT LatB-treated Tg(kdrl:nlsEGFP)zf109 embryos at the indicated stages. Arrowheads indicate patterned pErk-positive tip cells. Tip cell patterning is delayed in LatB-treated embryos. N = 4 (one experiment shown).

Left subpanel depicts quantified pErk staining as a function of distance along anterior–posterior axis, starting from the first patch of bright pErk staining for control and LatB embryos at 22 hpf. Right subpanel depicts spatial frequency spectra resulting from Lomb-Scargle analysis of quantified 22 hpf pErk stainings of control and LatB. All error bars show standard error of the mean. (Online version in colour.)

To provide the first experimental validation of these predictions in vivo we analysed tip selection in the dorsal aorta (DA) of zebrafish embryos. We chose developing zebrafish in particular because they are (i) easily amenable and well-validated for live-imaging angiogenesis being transparent during development [71] and (ii) sprouting of ISVs from the DA is a highly stereotypical process with simple trackable branch initiations arising from a single pre-existing DA vessel [47], very similar to the simulation set-up used above.

First, using high spatiotemporal resolution imaging of emerging tip cells in live zebrafish embryos, we confirmed our time-ordered argument to be correct, as filopodia were observed on ECs residing in the DA at timepoints prior to tip cell selection and ISV branching (figure 2d). We then used low-dose latrunculin B (LatB) from these early timepoints to specifically disrupt filopodia extension by perturbing F-actin nucleation at filopodia (but not other actin nucleation sites) [27,44]. A dose-response study previously found that low doses of LatB, within the range of 0.08 and 0.1 µg ml−1, specifically disrupt filopodia in zebrafish ISVs, without impacting cortical actin, cell shape changes or causing the myocardial embryonic lethal issues that arise with higher LatB doses [44].

This severely perturbed the speed of tip–stalk decision-making. Unusually, it also extended the otherwise stereotypical time window for tip cell selection (albeit by only 3 h; figure 2e). The extended time window was akin to observations in Dll4-deficient embryos, suggesting that the selection window is determined largely by filopodia and Notch, but not by Vegfr signalling [20]. Moreover, quantification of the number of ECs residing in the DA that are selected to branch revealed that a loss of filopodia severely restricts the number of tip cells selected to enter branching ISVs (figure 2e). Hence, these observations were consistent with a key role for filopodia-mediated temporal control of the collective decision-making process.

Next, we wanted to assess whether the reduced rates of tip cell selection observed in the absence of filopodia (figure 2e) were associated with defects in lateral inhibition and establishment of the salt and pepper patterning of Vegfr-active tip cells. To observe tip cell patterning, zebrafish embryos were immunostained for pErk, a highly-specific readout for active Vegfr in EC tip cells [20,27,72], at the indicated time points during tip cell selection (figure 2f). In control embryos, Vegfr-active ECs were abundant and heterogeneously spread throughout the dorsal aorta during early phases of tip cell selection (up to 21 hpf). This broad and random distribution then rapidly resolved to a refined, highly regular salt and pepper pattern of Vegfr-active ECs by 22 hpf, indicating robust tip cell patterning coincident with initiation of ISV branching, as previously described [73]. Consequently, tip cells that initiated ISV branching were consistently pErk-positive (figure 2f).

In contrast, the tip cell patterning process was delayed in LatB-treated embryos lacking filopodia (figure 2f). In particular, the earliest patterning step whereby ECs exhibited heterogeneous Vegfr activity throughout the DA was prolonged by up to 3 h in LatB-treated embryos, with non-patterned clusters of pErk-positive cells evident until 24 hpf and longer. Importantly, this perturbation of tip cell patterning was not due to an overall embryonic delay (data not shown).

To characterize the generality of this effect, we quantified the pErk pixel values within a rectangular region along the DA in ImageJ. The left side of the rectangle was aligned to the anterior side of the first patch of bright pErk staining seen in the DA for all of our data at 22 hpf (our highest data timepoint, N = 5, control n = 46, LatB n = 43). We then created a plot profile of the rectangle in ImageJ to generate a list of pErk pixel brightness values along the anterior-posterior axis. When we averaged the values of the pErk brightness profiles over all embryos, we found that the control had a more spatially periodic appearance of pErk brightness along the DA (figure 2f, left subpanel). This is descriptive of the fact that ECs in the DA at 22 hpf already have regular differential activation patterns corresponding to ISV locations across the fish in the control group. In the LatB condition, there were still some elements of periodicity, but much of it is lost when averaging across fish and experiments. This suggests that LatB impinges on the ability of the cells to reliably create a differential pattern in Vegfr activations.